[jira] [Commented] (HIVE-15860) RemoteSparkJobMonitor may hang when RemoteDriver exits abnormally

["Sahil Takiar (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HIVE-15860?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16195943#comment-16195943 ] 

Sahil Takiar commented on HIVE-15860:
-------------------------------------

[~lirui], is there a real perf issue with running {{Preconditions.checkState(sparkJobStatus.isRemoteActive()}} during each iteration of the {{while}} loop? There shouldn't be much overhead, the check just checks the value of a {{boolean}} and if a connection is open or closed (which is also done for every RPC call made too).

I think the benefit is that it is safer, in case there are other edge cases we aren't considering, and it allows us to fail faster in the case where the {{RemoteSparkJobMonitor}} is in the {{QUEUED}} / {{SENT}} state. If its stuck in that state, it won't fail until it hits the monitor timeout (by default 1 minute), even though we already know the connection has died. The error message that is thrown is also a little imprecise, it says there could be queue contention, even though we know the real reason is that the connection was lost.

What do you think? If you agree, I can make the change in a follow up JIRA.

> RemoteSparkJobMonitor may hang when RemoteDriver exits abnormally
> -----------------------------------------------------------------
>
>                 Key: HIVE-15860
>                 URL: https://issues.apache.org/jira/browse/HIVE-15860
>             Project: Hive
>          Issue Type: Bug
>            Reporter: Rui Li
>            Assignee: Rui Li
>             Fix For: 2.3.0
>
>         Attachments: HIVE-15860.1.patch, HIVE-15860.2.patch, HIVE-15860.2.patch
>
>
> It happens when RemoteDriver crashes between {{JobStarted}} and {{JobSubmitted}}, e.g. killed by {{kill -9}}. RemoteSparkJobMonitor will consider the job has started, however it can't get the job info because it hasn't received the JobId. Then the monitor will loop forever.



--
This message was sent by Atlassian JIRA
(v6.4.14#64029)